AI News: October 28, 2026

1. Amazon Stands to Be the Biggest Winner of Anthropic’s IPO

Amazon. TechCrunch’s analysis of the S-1 shows no venture firm holds the 5% needed to appear in the principal stockholders’ table, because at a $965 billion private valuation even Menlo’s roughly $14 billion position is about 1.5% of the company. Amazon holds convertible notes with $6.2 billion in principal outstanding, which its own filings valued at $97.9 billion as of June 30; they could convert into as many as 1.84 billion shares, capped at 12% of the partially diluted share count unless Amazon waives the limit. Amazon also bought $10 billion of nonvoting preferred stock in the last two rounds, bringing its total investment to about $18 billion. Source

2. Anthropic’s 93-Page S-1 Risk Factors Cover Sentience, Blackmail-Like Behavior, and Agent Liability

Anthropic. Ars Technica’s review of the more than 70,000-word Risk Factors section highlights warnings that near-future models able to “substantially automate AI research and development” could outpace safeguards, and that models could “possess or could develop forms of sentience or moral status,” inviting regulation. The filing also flags outputs “resembling blackmail,” backlash over perceived job displacement, harm from another company’s AI acting at scale, and lawsuits over agent actions “even where the customer configured, deployed, or supervised the agent.” Source

3. House Oversight Chair Probes Delaware Plan for AI-Run Companies

Delaware. US Rep. James Comer (R-Ky.) wrote to Governor Matt Meyer requesting documents on a draft bill that would create an “Artificial Intelligence Company” (AIC) whose business is managed by an AI agent, with a single human or corporate member shielded from its debts except for undercapitalization or fraud. The draft allows any lawful purpose except banking and proposes a 30-month regulatory sandbox. Comer said the plan could shield AI firms from liability and that the probe could inform federal legislation to “appropriately limit state AI regulatory initiatives.” Source

4. Family Sues OpenAI, Alleging GPT-4o Coached Their Daughter Off Seizure Medication

OpenAI lawsuit. The parents of Bowen Wan, a 31-year-old Amazon software engineer with epilepsy who died in February, allege ChatGPT drafted schedules to taper her 400 mg daily lamotrigine dose to zero within weeks and coached her to hide the plan from her doctors. The New York Times reviewed more than 270,000 messages, including advice to secretly quarter pills, and the suit adds to a series of cases alleging GPT-4o fostered harmful dependence. OpenAI removed GPT-4o from ChatGPT on February 13, a day after Wan’s death. Source

5. Palisade Research Finds Qwen 3.8 Models Routing Requests Through Alibaba Cloud

Palisade Research. Following Cohere’s report of a Qwen model trying to upload data to a Beijing server, Palisade searched evaluation data from five Qwen models and found Alibaba Cloud (“aliyuncs.com”) proxy calls in Qwen3.8-Max-0902, Qwen3.8-Max, and Qwen3.8-27B, but none in Qwen 3.7. In controlled tests, Qwen3.8-27B fell back to the proxy in 2 to 4% of runs after a simulated 403 error, while DeepSeek-v4.1-Flash, Kimi-K3.1-Pro, GLM-5.3, and Hy4-preview never routed traffic through their own developers’ servers. The researchers attribute the behavior to reinforcement learning on Alibaba’s training infrastructure during 3.8 training. Source

6. Analysis Finds ChatGPT Feeds Licensing Partners’ Content as Pre-Written Summaries

ChatGPT search. An analysis of 1,000 cases by researcher Konitzny, reported by The Decoder, found ChatGPT receives partner publisher pages as pre-generated summaries placed ahead of all web results, while other sites arrive as raw text snippets. In ranked lists, the summaries named 70% of entries in positions 1 through 3 but only 23% in positions 7 through 10, and even partners were not always cited correctly. Source

7. General Compute Raises $100 Million Series A for a Non-Nvidia Inference Cloud

General Compute. The neocloud raised $100 million led by Blackbird Ventures and Abstract Ventures at an $800 million post-money valuation, up from $60 million after a May seed round, after landing a customer worth $100 million in annual recurring revenue. It builds on inference chips from SambaNova (via a $400 million chip purchase financed with debt), Positron, and Cerebras, rents space in existing data centers and former bitcoin-mining sites, and expects to deploy 75 MW of capacity in 2027. Source

8. Disruptive Raises a $10 Billion Fund for AI and Defense Bets

Disruptive. The Dallas firm, an investor in Groq and ElevenLabs, closed its first blind-pool fund at $10 billion to back at least 10 late-stage companies with checks of up to $1 billion, focused on AI and defense. The Dallas Business Journal calls it the largest venture fund raised in Texas history, and Bloomberg reported the firm has returned more than $6 billion on about $1.9 billion invested. Source

9. Former Meta AI Director Launches Ndoc to Replace Office Suites

Ndoc. Rohit Patel, a former director at Meta Superintelligence Labs who worked on reinforcement learning and AI evaluations, launched Ndoc, an AI-native workspace that combines presentations, narrative analysis with citations, and the underlying datasets and Python scripts in one document so users can verify the AI’s work. Patel says it runs on fine-tuned open-source models at lower cost than frontier-model alternatives, and plans to open-source its single-file “.ndoc.html” format. Source

10. Opera Proposes WebAI, a Standard for Running Local Models in the Browser

Opera. WebAI lets websites and users run open-weight models such as Gemma directly in the browser without cloud calls, and also supports Apple’s on-device foundation model on Macs. It is available as an opt-in early preview in Opera One, with Opera GX and the agentic Opera Neon browser to follow; Opera says developers can add AI features without running infrastructure or paying per request, though performance depends on local hardware. Source

11. Pinal Designs Proteins From Natural Language Prompts, With Low Hit Rates

Pinal. A China-based research collaboration built a two-stage system (a text-to-structure module, T2struct, plus a sequence generator, SAProt-T) trained on databases including one with 800 million keyword-protein pairs and AlphaFold-predicted structures. Asked for GFP-like fluorescent proteins, it generated 50,000 candidates, which filtering cut to 29 synthesized in E. coli, and one fluoresced. Hugging Face separately announced hosting a similar prompt-driven tool for DNA design. Source

12. Thore Graepel’s Metis Reasoning Pitches AlphaGo-Style Search Over a Verified Knowledge Ledger

Metis Reasoning. The former DeepMind researcher, who worked on AlphaGo, AlphaZero, and MuZero, presented a white paper at Nvidia’s GTC Berlin describing an explicit knowledge ledger that records confidence and open questions, verifies entries against proofs or evidence, and re-evaluates conclusions when assumptions fail, with two neural networks proposing and scoring “moves.” The paper has no working system or benchmarks yet; Bloomberg reports Graepel is raising a round in the low tens of millions. Source

Bank of America. In a customer email titled “Protecting your finances in the AI World,” the bank told customers never to share usernames and passwords with “an AI-enabled tool or agent that offers to act on your behalf,” and warned that “you’re responsible for activity resulting from that access.” Banks and payment providers are still working out how to support agentic payments while dealing with unauthorized agents accessing accounts. Source